AURANET: Recurrent Alternate Update Training for Spatiotemporal Prediction with Incomplete Observations

Spatiotemporal prediction using data fusion from multiple sources presents significant challenges, especially without proper equipment or resources causing incomplete observations or missing data in the prediction hyperspace. Statistical interpolation methods and conventional machine learning models struggle to capture the true data distribution in highly sparse spatiotemporal settings. To address these limitations, we propose AURANET (Alternate Update with Recurrent Approach for Neural nETworks)—a novel training framework that alternately updates the model trainable parameters and the corresponding predictions iteratively until a convergence criterion is met for the available observations. AURANET operates in two major steps—(1) with sparse observations as input, it generates a recurrent chain-of-prediction to gap-fill a target spatiotemporal data cube with fixed network parameters (i.e., weights and biases), and (2) at the end of the prediction chain, it updates these network parameters only for the available ground truth observations, i.e., use a subset of the predictions to train the neural network at each training step. In this paper, AURANET is implemented to gapfill and predict daily spatial plant evapotranspiration (ET) data in Montana potato fields, where complete spatial observations are available only on certain days of the growing season, and complete temporal observations are available only for one specific location in the field. The corresponding predictions show a validation MSE of 1.36 mm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> d<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-2</sup> on the 2023 growing season and a test error of -1.47 mm d<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">-1</sup> on the 2024 growing season. The AURANET framework extends beyond ET prediction and can be applied to various domains requiring recurrent learning structure with incomplete data.